Leonard J. Trejo

Brain-computer interfaces for 1-D and 2-D cursor control: designs using volitional control of the EEG spectrum or steady-state visual evoked potentials

We have developed and tested two electroencephalogram (EEG)-based brain-computer interfaces (BCI) for users to control a cursor on a computer display. Our system uses an adaptive algorithm, based on kernel partial least squares classification (KPLS), to associate patterns in multichannel EEG frequency spectra with cursor controls. Our first BCI, Target Practice, is a system for one-dimensional device control, in which participants use biofeedback to learn voluntary control of their EEG spectra. Target Practice uses a KPLS classifier to map power spectra of 62-electrode EEG signals to rightward or leftward position of a moving cursor on a computer display. Three subjects learned to control motion of a cursor on a video display in multiple blocks of 60 trials over periods of up to six weeks. The best subjects average skill in correct selection of the cursor direction grew from 58% to 88% after 13 training sessions. Target Practice also implements online control of two artifact sources: 1) removal of ocular artifact by linear subtraction of wavelet-smoothed vertical and horizontal electrooculograms (EOG) signals, 2) control of muscle artifact by inhibition of BCI training during periods of relatively high power in the 40-64 Hz band. The second BCI, Think Pointer, is a system for two-dimensional cursor control. Steady-state visual evoked potentials (SSVEP) are triggered by four flickering checkerboard stimuli located in narrow strips at each edge of the display. The user attends to one of the four beacons to initiate motion in the desired direction. The SSVEP signals are recorded from 12 electrodes located over the occipital region. A KPLS classifier is individually calibrated to map multichannel frequency bands of the SSVEP signals to right-left or up-down motion of a cursor on a computer display. The display stops moving when the user attends to a central fixation point. As for Target Practice, Think Pointer also implements wavelet-based online removal of ocular artifact; however, in Think Pointer muscle artifact is controlled via adaptive normalization of the SSVEP. Training of the classifier requires about 3 min. We have tested our system in real-time operation in three human subjects. Across subjects and sessions, control accuracy ranged from 80% to 100% correct with lags of 1-5 s for movement initiation and turning. We have also developed a realistic demonstration of our system for control of a moving map display (http://ti.arc.nasa.gov/).

Kernel PCA for Feature Extraction and De-Noising in Nonlinear Regression

In this paper, we propose the application of the Kernel Principal Component Analysis (PCA) technique for feature selection in a high-dimensional feature space, where input variables are mapped by a Gaussian kernel. The extracted features are employed in the regression problems of chaotic Mackey–Glass time-series prediction in a noisy environment and estimating human signal detection performance from brain event-related potentials elicited by task relevant signals. We compared results obtained using either Kernel PCA or linear PCA as data preprocessing steps. On the human signal detection task, we report the superiority of Kernel PCA feature extraction over linear PCA. Similar to linear PCA, we demonstrate de-noising of the original data by the appropriate selection of various nonlinear principal components. The theoretical relation and experimental comparison of Kernel Principal Components Regression, Kernel Ridge Regression and ε-insensitive Support Vector Regression is also provided.

EEG-based estimation of mental fatigue: convergent evidence for a three-state model

Two new computational models show that the EEG distinguishes three distinct mental states ranging from alert to fatigue. State 1 indicates heightened alertness and is frequently present during the first few minutes of time on task. State 2 indicates normal alertness, often following and lasting longer than State 1. State 3 indicates fatigue, usually following State 2, but sometimes alternating with State 1 and State 2. Thirty-channel EEGs were recorded from 16 subjects who performed up to 180 min of nonstop computer-based mental arithmetic. Alert or fatigued states were independently confirmed with measures of subjects performance and pre- or post-task mood. We found convergent evidence for a three-state model of fatigue using Bayesian analysis of two different types of EEG features, both computed for single 13-s EEG epochs: 1) kernel partial least squares scores representing composite multichannel power spectra; 2) amplitude and frequency parameters of multiple single-channel autoregressive models.

Feature extraction of event-related potentials using wavelets: an application to human performance monitoring.

This report describes the development and evaluation of mathematical models for predicting human performance from discrete wavelet transforms (DWT) of event-related potentials (ERP) elicited by task-relevant stimuli. The DWT was compared to principal components analysis (PCA) for representation of ERPs in linear regression and neural network models developed to predict a composite measure of human signal detection performance. Linear regression models based on coefficients of the decimated DWT predicted signal detection performance with half as many free parameters as comparable models based on PCA scores. In addition, the DWT-based models were more resistant to model degradation due to over-fitting than PCA-based models. Feed-forward neural networks were trained using the backpropagation algorithm to predict signal detection performance based on raw ERPs, PCA scores, or high-power coefficients of the DWT. Neural networks based on high-power DWT coefficients trained with fewer iterations, generalized to new data better, and were more resistant to overfitting than networks based on raw ERPs. Networks based on PCA scores did not generalize to new data as well as either the DWT network or the raw ERP network. The results show that wavelet expansions represent the ERP efficiently and extract behaviorally important features for use in linear regression or neural network models of human performance. The efficiency of the DWT is discussed in terms of its decorrelation and energy compaction properties. In addition, the DWT models provided evidence that a pattern of low-frequency activity (1 to 3.5 Hz) occurring at specific times and scalp locations is a reliable correlate of human signal detection performance.

Measures and Models for Predicting Cognitive Fatigue

We measured multichannel EEG spectra during a continuous mental arithmetic task and created statistical learning models of cognitive fatigue for single subjects. Sixteen subjects (4 F, 18-38 y) viewed 4-digit problems on a computer, solved the problems, and pressed keys to respond (inter-trial interval = 1 s). Subjects performed until either they felt exhausted or three hours had elapsed. Pre- and post-task measures of mood (Activation Deactivation Adjective Checklist, Visual Analogue Mood Scale) confirmed that fatigue increased and energy decreased over time. We examined response times (RT); amplitudes of ERP components N1, P2, and P300, readiness potentials; and power of frontal theta and parietal alpha rhythms for change as a function of time. Mean RT rose from 6.7 s to 7.9 s over time. After controlling for or rejecting sources of artifact such as EOG, EMG, motion, bad electrodes, and electrical interference, we found that frontal theta power rose by 29% and alpha power rose by 44% over the course of the task. We used 30-channel EEG frequency spectra to model the effects of time in single subjects using a kernel partial least squares (KPLS) classifier. We classified 13-s long EEG segments as being from the first or last 15 minutes of the task, using random sub-samples of each class. Test set accuracies ranged from 91% to 100% correct. We conclude that a KPLS classifier of multichannel spectral measures provides a highly accurate model of EEG-fatigue relationships and is suitable for on-line applications to neurological monitoring.

Kernel PCA Feature Extraction of Event-Related Potentials for Human Signal Detection Performance

In this paper, we propose the application of the Kernel PCA technique for feature selection in high-dimensional feature space where input variables are mapped by a Gaussian kernel. The extracted features are employed in the regression problem of estimating human signal detection performance from brain event-related potentials elicited by task relevant signals. We report the superiority of Kernel PCA for feature extraction over linear PCA.

Estimation of Human Signal Detection Performance from Event-Related Potentials Using Feed-Forward Neural Network Model

We compared linear and neural network models for estimating human signal detection performance from event-related potentials (ERP) elicited by task-relevant stimuli. Data consisted of ERPs and performance measures from five trained operators who monitored a radar display and detected and classified visual symbols at three contrast levels. The performance measure (PF1) was a composite of accuracy, speed, and confidence of classification responses. The ERPs, which were elicited by the symbols, were represented in the interval 0-1500 ms post-stimulus at three midline electrodes (Fz, Cz, Pz) using either principal component analysis (PCA) factors or coefficients of autoregressive (AR) models. We constructed individual models of PF1 from both PCA and AR representations using either linear regression or radial basis function (RBF) networks. Applying the normalized mean square error of approximation as a criterion, we found that the PCA representation was superior to AR and that RBF networks estimated PF1 much more accurately than linear regression. This suggests that nonlinear methods combined with suitable ERP feature extraction can provide more accurate and reliable estimates of display-monitoring performance than linear models.